A simple market-basket recommender on a Thai B2B platform sharply increased engagement and sales: orders including recommended items rose 64% and recommended-item sales grew 34%, while product exposure expanded more than tenfold; however, the study lacks a control group to confirm causality.

Main Finding

A market-basket-analysis recommendation system based on association rule mining (Apriori) was integrated into a Thai B2B e-commerce platform for traditional wholesalers and retailers. After deployment the system markedly increased engagement and commercial outcomes: the share of orders containing recommended items rose by 64%, the diversity of recommended products increased 1,060%, purchases of recommended products by value grew 34%, and user satisfaction was high (Net Promoter Score = 55%). The study shows that a low-complexity, item-relationship approach can materially boost product exposure and revenue in data‑limited traditional trade settings.

Key Points

• Problem: Traditional B2B/SME trade in Thailand lags modern e-commerce; fragmented actors and sparse user behavioral data make mainstream recommendation methods (e.g., collaborative filtering) less suitable.
• Approach: Use association rule mining (market basket analysis) to produce item-to-item recommendations (X → Y) rather than relying on top-seller lists or user-profile heavy methods.
• Algorithm: Apriori algorithm for frequent itemset discovery and rule generation; rules evaluated with support, confidence and lift.
• Integration: The recommendation engine was integrated into an existing B2B e-commerce platform serving 1,168 retailers and 45 wholesalers.
• Evaluation metrics: order behavior (share of orders including recommended products), product diversity (items surfaced), sales value of recommended items, and retailer satisfaction (NPS).
• Main quantitative outcomes: +64% share of orders with recommended items; +1,060% in recommended-product diversity; +34% in monetary value of recommended-product purchases; NPS = 55%.
• Practical contributions: demonstrates feasibility of ARM-based recommender in resource-constrained environments and documents operational observations (e.g., parameter and usage impacts).
• Identified operational improvements: automating rule updates and improving usability as next steps.

Data & Methods

• Data source: Transactional dataset from a B2B e-commerce platform (retailers and wholesalers) covering January 2021 – March 2022. (Transaction count not specified in the provided excerpt.)
• Sample: Platform ecosystem with 1,168 retailers and 45 wholesalers.
• Preprocessing: Market-basket formation from transactional records (itemsets per order); standard cleaning and aggregation implied.
• Algorithmic method: Apriori algorithm to discover frequent itemsets and generate association rules that meet minimum support and confidence thresholds; rules screened/selected for recommendation use (lift also considered for rule strength).
• Deployment: Generated rules embedded into the platform recommendation pipeline (item-to-item suggestions presented to retailers).
• Evaluation design: Before/after observational assessment of platform metrics (proportion of orders with recommendations, diversity of recommended items, value of purchases from recommended items) plus a retailer satisfaction survey (NPS). No randomized control group reported.
• Limitations in methods reported or implied: reliance on transactional co-occurrence (no individualized profiles), potential need for automated and periodic re-mining of rules, and absence of an experimental (causal) identification strategy in the reported evaluation.

Implications for AI Economics

• Low-cost recommendation tech can raise revenues in fragmented SME markets: ARM-based recommenders are relatively simple to implement and interpret, making them well-suited for firms/platforms with limited data engineering resources and scarce per-user interaction history.
• Long-tail and product-diversity effects: A 1,060% increase in recommended-product diversity suggests MBA can surface rarely purchased items, shifting sales composition away from top-sellers and potentially improving welfare for niche suppliers and reducing over-reliance on a few SKUs.
• Platform competitiveness and digital transformation: Effective recommender features can help traditional-trade platforms compete with modern trade and marketplaces by increasing retailer engagement and satisfaction, contributing to broader digital adoption among SMEs.
• Operational economics: The results imply positive ROI potential for wholesalers/platforms investing in simple ARM recommenders—benefits include higher order attachment rates and larger recommended-item spend. However, platform costs (compute, rule refresh cadence, UX) and integration costs must be weighed.
• Market-structure and strategic considerations: Increased cross-selling may alter ordering patterns, inventory turnover, and bargaining dynamics between wholesalers and retailers. Platforms may gain more value capture if they can sustain higher engagement and broadened assortment demand.
• Research and policy directions:
  • Causal evaluation: Future work should use randomized trials or quasi-experimental methods to estimate causal effect sizes on retailer profits, ordering frequency, and inventory dynamics.
  • Comparative benchmarking: Compare ARM recommenders with collaborative, content-based, and hybrid methods in traditional B2B contexts to identify trade-offs in accuracy, diversity, and implementation cost.
  • Dynamic / automated pipelines: Study optimal rule-refresh frequencies and online learning to handle seasonality and evolving assortments.
  • Welfare and competition analysis: Examine whether increased exposure of long-tail items benefits small suppliers and retailers or leads to cannibalization of existing sellers; analyze impacts on market concentration.
  • Scalability and generalizability: Test the approach across different geographies, product categories, and platform sizes to map where ARM provides the largest economic returns.
• Cautionary notes: Observational before/after results are promising but not definitive proof of causation; platform-level adoption and sustaining gains may depend on UX, incentives, and rule maintenance.